Int. J. Fin. Acco. Eco. Stu. Vol. 2 / No.

1 / Winter 2012

Investigating Financial Crisis Prediction Power using Neural Network and

Non-Linear Genetic Algorithm

Receipt: 19, 6 , 2012

Acceptance: 25, 7 , 2012

Zahra Poorzamani
Assistant Professor, Department of Accounting, Central Tehran Branch, Islamic Azad University, Tehran, Iran
Corresponding Author,
zpoorzamani@yahoo.com

Hassan Kalantari
M.Sc. student in Accounting at Islamic Azad University, Central Tehran Branch
kalantari.hassan@yahoo.com

Abstract
Bankruptcy is an event with strong impacts on management, shareholders, employees,
creditors, customers and other stakeholders, so as bankruptcy challenges the country both
socially and economically. Therefore, correct prediction of bankruptcy is of high importance
in the financial world. This research intends to investigate financial crisis prediction power
using models based on Neural Networks and to compare it with Non-Linear Genetic
Algorithm. Based on the available information and statistics of the listed companies on
Tehran Stock Exchange (TSE) during 1997-2010, from among these companies subjected to
article 141 of the Commercial Law, 72 firms, and from among other firms, 72 firms were
selected.
Results of McNemar Test for Non-Linear Genetic Algorithm and Neural Network
indicated that although prediction accuracy of Non-Linear Genetic Algorithm (90%) was
greater than that of Neural Network (70%), yet this difference was not statistically
significant
Keywords: bankruptcy prediction, Non-Linear Genetic Algorithm, Neural Network.

10 /

Investigating Financial Crisis Prediction Power using Neural Network

1- Introduction
There are numerous factors that affect
the bankruptcy phenomenon. High interest
rate and heavy debts are among the factors
which negatively affect the firms financial
state. In addition, prior research suggests
that newly founded private firms and
smaller companies, respectively, are more
vulnerable compared to established and
large companies (Dun and Bradstreet,
1980).
Competition intensification at industry
level has led to bankruptcy of many firms
and their removal from the competition
field. This has given rise to some concerns
among shareholders, managers, creditors
and in general the whole society. Investors
by estimation of financial crises and
bankruptcy of firms try to prevent loss of
their capital. If management of business
unit is timely informed of bankruptcy risk,
it can take preventive actions. Creditors are
very sensitive about loss of their principal
amount and interest in the granted loans
and credits to potential and current
customers, and since it imposes heavy
economic and social bankruptcy costs on
the society, it is also interesting from
macroeconomic point of view, because the
lost resources in the distressed economic
unit could have been used for other
profitable opportunities. Given the
importance of bankruptcy, all people and
stakeholders are interested in bankruptcy
prediction before its actual occurrence.
Financial
crisis
or
bankruptcy
prediction using historical financial data is
well known. Although the first effort in
this relation dates back to 1930, but from
1966 and following the research carried
out by Beaver on this topic it took a more
serious form (Dimitras et al, 1996). Beaver
(1966) is one of the first researchers who
investigated prediction of financial crisis
or bankruptcy and is regarded as one of the

leading academic researchers in this field.

After him, Altman (1968) using advanced
statistical techniques succeeded in
achieving significant results. Today,
numerous prediction models are introduced
by researchers. These models, according to
the model construction method, number of
models variables, definition of bankrupt
firms and other firms, are classified into
various groups, such as traditional model
versus Artificial Intelligence models, or
univariate models versus multivariate
models.
This research, in addition to setting the
prior research in financial crisis prediction
as the departure point which is based on
data of one or several years, intends to
construct and design a financial crisis
prediction model based on financial
variables during the understudy years
using the Artificial Intelligence and
Genetic Algorithm corresponding to
Iranian economic condition.
2.Literature of Review and research
background
Large scale financial crisis following
financial distress and bankruptcy of firms
operating in an economy by extension to
crises at regional level may have
international impacts. Financial distress in
Oxford Dictionary is defined as hardship,
pain, grief and lack of monetary resources
and destitute. In financial literature, a
variety of definitions have been offered for
financial distress.
In this research, bankrupt firm is
defined based on article 141 of the
Commercial Law according to which a
bankrupt company has an accumulated loss
equal to 50% of the firms capital.
However, article 141 of the Commercial
Law does suggest an immediate
dissolution or liquidation of the firm, and

Vol.2 / No.1 / Winter 2012

International Journal of Finance, Accounting and Economics Studies / 11

only upon request of stakeholders the
firms activities are suspended.
Identification of exact reason(s) of
bankruptcy and financial problems in
every particular case is not easy. In most
cases, a multiplicity of reasons altogether
leads to financial crisis or bankruptcy. But,
in general, the main causes of bankruptcy
are the mentioned financial and economic
problems.
In some cases, causes of financial crisis
or
bankruptcy are identified by
examination of financial statements. The
accountants who are experienced in
analysis of the firms financial state can
easily detect causes of financial crisis or
bankruptcy. However, at times, some
indices indicate favorable financial flow in
a business unit for a relatively short period,
concealing financial crisis or bankruptcy
from the eyes of accountants.
Newton (1998) has divided the stages
of firms financial state worsening into
incubation
period,
cash
shortage,
insolvency for payment of financial and
commercial liabilities, insolvency for
payment of total debt, and eventually
financial crisis and bankruptcy (figure 1.2).
Although most bankruptcies go through
these stages, but some companies may
without undergoing all these stages reach
financial crisis and complete bankruptcy.
The state of business unit does not
abruptly enter financial crisis or
bankruptcy stage. In incubation period,
there may be one or more unfavorable
situation for the business unit, without
immediately being
detectable. For
example, change in production demand,
steady increase in overhead costs,
abolishment of production methods, etc are
among such factors. Often, economic
losses occur during incubation period
which reduce return on assets (ROA). If
the problem is discovered at this stage, it

would be ideal for the company,

particularly since at this stage easier
solutions are effective, while at later stages
they may be of no use. The third point is
that if the problem is discovered and
resolved at this stage, public confidence in
the firm will be undermined. Resolving the
problem at later stages weakens public
trust in the firm and as a consequence of
which access to funds will be more
difficult and the firm may be forced to
refuse profitable investment projects.
The stage of cash deficit starts when the
business unit for the first time does not
have enough cash available to meet its
current liabilities or future needs, although
it may have much more physical asset
relative to its current needs and
profitability records. The point is that the
assets cannot sufficiently be liquidated and
the capital is in fact locked up.
At the stage of financial and
commercial insolvency, the firm is still
able to obtain sufficient funds from
consumption channels, management is in a
position to takes suitable measures such as
use of financial and commercial
specialists, credit granting committee, and
financing methods restructuring. Taking
these actions, the firm is still able to
identify and resolve the financial problem.
At bankruptcy stage, the firm moves
towards extinction. At this stage, total debt
exceeds total asset value, and the firm can
no longer avoid suspension of its
operation, followed by rush of creditors
and other stakeholders to collect their
claims. At any rate, it should be noticed
that bankruptcy is defined according to the
law and should be considered from this
perspective.
The extracted ratios and indices from
financial statements have been always
considered as signals of firms current and
future financial state (Galvao, 2004). Use

Vol.2 / No.1 / Winter 2012

12 /

Investigating Financial Crisis Prediction Power using Neural Network

of financial ratios is the most common

technique in bankruptcy prediction.
Application of financial ratios to
bankruptcy prediction dates back to 1966
and Beavers studies which for a long time
has been regarded as the only possible
method (Haber, 2006). In recent years,
many critics have been made on prediction
power of financial ratios. However, Beaver
(2005) demonstrated that financial ratios
are still the powerful tools for bankruptcy
prediction.
2.1. Techniques employed in bankruptcy
prediction
Due to the significant economic, social
and political consequences imposed by
bankruptcy phenomenon on various groups
in the society, it has been always of special
interest with researchers. Researches on
bankruptcy prediction have taken a serious
form since 1960, so as today we are
witness of numerous and diverse
methodologies applied to to bankruptcy
prediction. This section addresses these
techniques
and
describes
their
characteristics features and limitations.
In most studies on bankruptcy
prediction, firms are basically classified
into two groups: 1. financially healthy
firms and 2. financially distressed firms.
Based on this classification, firms can be
classified by the dummy variable Y:
1

Yi = 0

Some researchers suggest definition of

more than two groups based on risk level,
but due to general acceptance of the 2category classification, it has become the
dominant approach (Dimitras et al, 1996).
Bankruptcy
prediction
techniques
according to their nature are classified into
three groups of (classical) statistical

approaches,
Artificial
Intelligence
techniques, and Theoretical models.
Data mining models and Artificial
Intelligence Techniques (AIT): AIT
performs tasks similar to humans
knowledge, intelligence and logic. In fact,
the AI is a system which learns and
improves performance of its problem
solving given the past experiences. AI
application in finance and particularly in
bankruptcy prediction does not have a long
record, yet due to its high efficiency and
being free from the existing restrictive
assumptions in statistical methods, it has
been widely accepted by the researchers.
These models are mainly focused on
signals of commercial failure, are generally
multivariate and the used variables in them
are derived from the information available
in the firms accounts. Intelligent
techniques are composed of neural
networks, genetic algorithms, hard sets,
Support Vector Machine, reasoning based
on Fuzzy and logic and issues. Many
studies have been carried out on
application of these techniques for
prediction of businesses failure among
which it can be referred to Etemadi,
Rostami and Farajzadeh Dehkordi (2009),
Huang, Tsai, Yen and Cheng (2008), Hung
and Chen (2009), Lin et al (2009), Min an
Jeong (2009), Min and Lee (2008), Ravi
and Pramodh (2008), Sun and Li (2008),
and Wu (2010).
For structuralization of Computer
systems neural networks, human learning
process and inference pattern are followed.
Architecture of neural networks in general
is consisted of three input layers including
input information, throughput (hidden)
layer, and output layer. Identification of
the best architecture for problem solving is
a complex and difficult task, and the best
architecture is obtained by trial and error.

Vol.2 / No.1 / Winter 2012

International Journal of Finance, Accounting and Economics Studies / 13

Neural Networks which are also known
as Artificial Neural Networks (ANN) are a
data mining technique used to solve many
problems. Among advantages of Neural
Networks relative to other methods,
according to the rules observed in data
mining and artificial intelligence systems,
it can be referred to the following ones:
Since Neural Networks do not need a
knowledge
base
for
being
structuralized, their use in problems
about which there is little knowledge
is useful.
Processing in NN can be performed at
high speed and accuracy relative to
traditional methods, because these
networks simultaneously examine all
the information existing in one
problem and the processing units or
the neurons function along each other
(Back et al, 1996).
In these networks, the type of data
distribution
or
communication
structure of the existing variables does
not require to be considered as the
basic assumption (Wu et al, 2006). If
the input data are incomplete and
disturbed or have a high correlation
with each other, or have not been
already observed, traditional systems
are hardly able to extract rules and
patterns, but in the same conditions,
neural networks provide reasonable
answer, and after learning and
adaptation, they will be able to
generalize the results to similar
instances.
Low energy consumption, error
tolerance, and high learning and
adaptation capability are other
advantages of this method.
However,
NNs
have
some
disadvantages as well. Their main
limitation is lack of a definitive method for

specification of optimal architecture. To

design an NN model in solving a problem
of classification type, there is no
systematic principle and method, as a
result, the best network topology is
specified by trial and error. Many factors
such as hidden layers, number of neurons
in hidden layers, data normalization and
learning algorithm can affect the networks
performance. In defining networks
architecture, one has to take this fact into
account that a greater number of layers
leads to more complexity of the network
and to a problem called over-fitting and
non-usability of new data. In general, with
increase in number of the middle neurons,
the networks power in identification of
existing complexities increases, but this
may reduce the networks generalizability.
In other words, if number of neurons in the
middle layer is too large, the network
memorizes in place of learning. Another
shortcoming of these networks concerns
their performance as a black box.
Comprehension and verification of the
mode these networks classify data and
quality of relationships in layers structure
is not possible for the user. NN method
does not specify significance of each
variable in the final classification, and the
assigned weights in this regard are not
interpretable (Fallahpour, 1383). From
among the researchers who employed NNs
for prediction of bankruptcy and financial
crisis in different countries it can be
referred to Coates and Fant (1992),
Serrano and Cinca (1996), and Shah and
Murtaza (2000).
By contemplation on bankruptcy
prediction models, it can be found that all
of them are somehow the heritage of
statistical techniques. For instance, AI
models generally use both univariate and
multivariate techniques so as they can be
considered as the children of mechanized

Vol.2 / No.1 / Winter 2012

14 /

Investigating Financial Crisis Prediction Power using Neural Network

statistical techniques. Similarly, theoretical

models often are derived from application
of an appropriate statistical technique and
are not directly derived from theoretical
principles.
Artificial Neural Networks (ANN) is
flexible and non-parametric modeling
tools. They are able to execute every
complex function with a satisfactory
accuracy. The first effort for use of ANNs
for bankruptcy prediction was made by
Adam and Sharda (1999).
Franco Varreto (1998) used Genetic
Algorithm for bankruptcy prediction. His
sample
included
500
companies,
comprising 236 bankrupt firms and 264
non-bankrupt firms. Results of this
research indicated a prediction accuracy of
93% one year ahead of the bankruptcy
event and a prediction accuracy of 91.6%
three years prior to bankruptcy. In
addition, in this research, by comparing
Genetic Algorithm with traditional
prediction models, it was voted for
superiority of genetic process, because
these models in addition to being free from
the restrictive assumptions, relative to
traditional methods have a higher
accuracy. In traditional models, with
increase of time horizon from bankruptcy
event, the models accuracy significantly
decreases, while this accuracy reduction is
far less with GA models. Among other
studies on this subject, it can be referred to
Shin and Lee (2002) and McKee and
Lensberg (2002).
Despite the large number of researches
on the issue of bankruptcy prediction, few
desirable results have been found (Plat &
Plat, 1990). The created models did not
come successfully out of the Test of
Robustness. This may be due to several
factors. One reason can be that the
statistical models tend to use Matched
Pairs of bankrupt and non-bankrupt firms.

Next, data of the variables specifying

cutoff points (thresholds) have been used
for distinction of bankrupt from nonbankrupt firms. The used data for
extraction of thresholds were year specific,
and were tried for correct identification of
firms in the hold-out sample (similar time
periods). These cutoffs have less efficiency
when used in different time periods
(including their use in next studies).
Another cause of the accuracy drop in
researches on bankruptcy prediction can be
sought in the selected variables for model
construction. Most studies select the used
variables in the model based on their
prevalence and desirability in the literature.
They often have root in the notion of
liquidity signal suggesting liquidity as a
synonym to financial solvency. Therefore,
these variables have been used in next
studies with the least manipulation and
modification.
Farajzadeh Dehkordi (2005), in his
master thesis, investigated bankruptcy
prediction modeling of the listed
companies on the stock exchange using
two models of Multiple Discriminant
Analysis (MDA) and Genetic Planning. To
construct the above models, first, he
prepared a full list of financial ratios (93
ratios) and after study of the ratios,
eventually, he extracted 42 financial ratios
for construction of the models, and using
Independent Samples Test (IST) of the two
societies he constructed the two intended
models. In fine, the Genetic Planning and
MDA models succeeded in correct
classification of the firms present in the
training set with an accuracy of 94% and
77%, and firms present in the hold-out set
with an accuracy of 90% and 73%,
respectively.
Kiarasi (2009) investigated two models
of Logistic Regression and Multivariate
Discriminant Analysis for prediction of the

Vol.2 / No.1 / Winter 2012

International Journal of Finance, Accounting and Economics Studies / 15

firms success or failure. He used 14
financial ratios and the results favored the
regression model over the Multivariate
Discriminant Analysis.
Saadatfar in his master thesis sought for
the best neural network structure for
prediction of firms bankruptcy. He used
three financial ratios of current ratio, gross
profit margin and net profit to current debt
to predict bankruptcy two years ahead of
its actual occurrence. His results indicated
superiority of the three-layer Neural
Networks (Perceptron model with a
structure made up of three neurons in the
first layer, nine neurons in the middle
layer, and one neuron in the output layer
with learning algorithm Error-Back
Propagation and cumulative learning
method and Sigmoid Activity Function) to
the four-layer Neural Networks.
Saee (2008) investigated application of
Support Vector Machine in prediction of
firms financial insolvency using financial
ratios. His research was mainly focused on
the use of Support Vector Machine in
prediction of firms insolvency. The results
obtained from this model were compared
with those obtained from the Logistic
Regression model and Support Vector
Machine was found superior to Logistic
Regression model.
Nouraddin (2010) constructed some
models for prediction of financial crisis
two years ahead of its occurrence. The
obtained results indicated superiority of the
Neural Networks model to other models.
3. Research Hypothesis
Given the purposes in this research, the
main hypothesis is as follows:
Prediction power of Neural Networks
(NN) models is greater than that of AIT
models which are based on internal
analysis (Non-linear Genetic Algorithm).

Since the financial crisis arises from

financial distress and bankruptcy of the
individual firms operating in the economy,
in this research, for detection of financial
crisis, it is made use of the firms
bankruptcy.
4. Statistical population and sample
The understudy statistical population in
this research includes all the listed firms on
TSE in the period 1997-2010. Based on the
available information and statistics in the
library of the Securities and Exchange
Organization (SEC), during this period, 72
firms have been subjected to article 141 of
the Commercial Law and hence, subjected
to the random selection. To match them
with the bankrupt firms in the same
number, non-bankrupt firms using random
sampling were selected. Due to the limited
number of the listed firms on the stock
exchange, it was not possible to compare
the firms in terms of industry in which they
operated. Since the firm size itself is
considered as a potential variable for
bankruptcy prediction, it is also dispensed
with comparison of the firms based on the
firm size. Therefore, in the sampling, the
non-bankrupt firms were matched with the
bankrupt firms only based on fiscal year.
5. Variable selection process
By accepting the assumption that
various financial ratios represent firms
financial state, they can be used as the
predicting variables of financial crisis or
bankruptcy. To determine suitable ratios
and indices for prediction of financial
crisis or bankruptcy, the research literature
has been thoroughly reviewed the result of
which was a list of 23 financial ratios used
in the prior research for prediction of
financial crisis or bankruptcy. A full list of
these ratios is provided in table 1.

Vol.2 / No.1 / Winter 2012

16 /

Investigating Financial Crisis Prediction Power using Neural Network

Table 1: Independent variables used in the research

Variable
X1
X2
X3
X4
X5
X6
X7
X8
X9
X10
X11
X12

Financial ratio
Working capital to equity
Working capital to sales
Working capital to total debt
Working capital to total asset
EBIT to equity
EBIT to sales
EBIT to total debt
EBIT to total asset
Equity to total debt
Equity to total asset
Sales to total debt
Sales to total asset

In selection of these variables, first, the

variables were refined according to the
theory in which variables with quite
similar effects were eliminated. For
example, although the financial ratios
total asset to total debt and total debt to
total asset are numerically different, in
fact, they represent an identical dimension
of firms state and simultaneous presence
of the two ratios only increases complexity
of the required calculations for selection of
the final variables. In the next stage, the
ratios and indices the calculation of which
was not possible based on the available
information on the stock exchange were
eliminated. Since use of all the 23 ratios
for construction of the model is not
possible and increases the models
complexity and reduces its efficiency, a
solution ought to be found in order to
reduce the number of the financial ratios in
a way that the models efficiency does not
suffer.
To select the predictive variables,
Stepwise Discriminant Analysis (SDA) is
used. SDA is one of the exploratory
elements of discirminant analysis which is
often used for the purpose of variable
selection. SDA technique, in the first step,
searches through the 23 ratios to determine

Variable
X13
X14
X15
X16
X17
X18
X19
X20
X21
X22
X23

Financial ratio
Total debt to accumulated profit or loss
Total debt to total asset
Accumulated profit or loss to total asset
Operational margin to sales
Financial cost to gross profit
Current asset to total asset
Sales to current asset
Current asset to current debt
Net profit to sales
Net profit to total asset
Current debt to total asset

the financial ratio with the greatest power

in distinction of bankrupt firms from the
non-bankrupt ones. To do this, SDA uses
the ratio of intra-group variance to
intergroup variance. So the financial ratio
with the least intra-group variance (within
each group of bankrupt and non-bankrupt)
and the greatest inter-group variance
(between two groups of bankrupt and nonbankrupt) has the greatest power in
distinction of the bankrupt from the nonbankrupt firms, because similarity in value
of a financial ratio in the group of bankrupt
firms, and on the other hand, the evident
difference of these values from the figures
of the non-bankrupt firms enhances
distinction power of this ratio and reduces
its errors.
After selection of the first variable,
SDA process is resumed for selection of
the second variable. At this stage, SDA
selects the financial ratio that next to the
first financial ratio has the greatest power
in distinction of firms. In fact, at this stage,
the joint effect of the two financial ratios in
dividing the groups of companies is
considered. This process is stopped when
none of the not-selected ratios meet the
specified condition of significance level
(i.e. 0.1). In other words, the made

Vol.2 / No.1 / Winter 2012

International Journal of Finance, Accounting and Economics Studies / 17

improvement by these financial ratios (notselected ones) in discriminating ability of
the selected ratios is insignificant. It should
be noted that in SDA process, because of
continuous evaluation of the sets
discriminating ability by F-Test, a financial
ratio may several times enter and exit the
set of the selected ratios. So a selected
financial ratio may be removed and
another financial ratio may come in its
place. Table 2 presents the obtained results
from the SDA process on 23 variables in
the previous stage. In this table, the 23
financial ratios together with their
significance degree in making distinction
between the firms as well as the final
variables are represented. Note that after
selection of the first financial ratio with the
highest significance level, the second
selected financial ratio is X14 which in
terms of significance is ranked 9 amidst 25
financial ratios. The reason for this
selection is the high correlation of the
significant financial ratios in second to

eighth row with the financial ratio X14 and

selection of these ratios does not increase
discriminating power of the set of the
selected variables. This also applies to
other selected financial ratios.
In addition, it should be noted that since
the specified significance level for
inclusion or removal of the variables is set
at 0.1, SDA process stops after selection of
the fifth financial ratio since other ratios
are not able to increase discriminating
power of the set of the selected variables at
least to the amount of 0.
Table 3 shows stages of SDA process.
As is seen in this table, to select the final
financial ratios, SDA process goes through
five stages and at each stage one financial
ratio is selected and added to the set of the
financial ratios. Decrease of Wilks
Lambda at each stage means increase of
discriminating power of the selected
variables.

Table 2: Matrix of SDA process structure

Significance
.243
.222
.221
-.157
-.102
-.044
.026

Stages
1
2
3

Code
X19
X23(a)
X15(a)
X13(a)
X17
X4(a)
X5(a)

Significance
-.437
.359
.344
.333
.318
.305
.270
.252

Code
X14
X8(a)
X11(a)
X9(a)
X20(a)
X18
X10(a)
X12(a)

Significance
.857
.832
.750
.745
.721
.636
.550
.477

Table 3: Summary of steps in variable selection process

Code
Change degree
F for exit
X16
1.000
71.858
X16
.973
76.804
X14
.973
6.212
X16
.804
35.601
X14
.925
8.740
X18
.766
6.281
X16
.771
26.502
X14
.826
12.246
X18
.724
8.552

Vol.2 / No.1 / Winter 2012

Code
X16
X1(a)
X3(a)
X21(a)
X2(a)
X22(a)
X6(a)
X7(a)

Wilks Lambda
.986
.697
.787
.681
.671
.731
.675
.661

18 /

Investigating Financial Crisis Prediction Power using Neural Network

Stages

Code
X19
X16
X14
X18
X19
X17

Change degree
.851
.738
.814
.714
.801
.708

Given the mentioned values, to construct

the model, SDA process selects 5 variables
from among the 23 candidate variables.
These 5 variables are ranked below
according to their discriminating power:
1) Ratio of operational margin to sales
(profitability): X16
2) Ratio of total debt to total asset
(solvency): X14
3) Ratio of current asset to total asset
(liquidity): X18
4) Ratio of sales to current asset
(efficiency): X19
5) Ratio of financial cost to gross profit
(interest coverage)
As is observed, each selected ratio
covers one important aspect of each the
firms
financial
state.
Statistical
significance of these ratios based on F-Test
is 0.857, -0.437, 0.305, 0.243, and -0.102,
respectively.
6. Construction of financial crisis or
bankruptcy prediction model using NonLinear Genetic Algorithm
The set of the understudy data which
includes 72 bankrupt and 72 non-bankrupt
firms has been randomly divided into two
groups of training set and hold-out set. The
training set which is used for construction
and training of the model includes 51
bankrupt firms and 53 non-bankrupt firms.
The hold-out set which includes 21
bankrupt firms and 19 non-bankrupt firms
is used to examine generalizability of the
obtained model and its external validity.
To execute Genetic Algorithm process and

F for exit
4.726
18.332
9.843
7.656
4.240
3.213

Wilks Lambda
.646
.711
.654
.641
.635
.617

to create financial crisis or bankruptcy

prediction
model,
GeneXproTools
software (version 4.0) is used. The cutoff
and mutation operators have been
determined at 0.6 and 0.06, respectively.
When the obtained result from Genetic
Algorithm for a company is greater than or
equal to 0.5 (threshold value), this
company is placed in the group of
bankrupt firms. On the contrary, when the
obtained value from the Genetic Planning
model is smaller than 0.5, the company is
placed in the group of non-bankrupt firms.
Comparison of the firms real group with
their predicted group, using Genetic
Algorithm, measures the models accuracy.
Diagram 1 presents the best model
obtained from the Non-Linear Genetic
Algorithm. This decision tree represents a
chromosome. Result of this tree for a firm
has to be compared with the threshold
value 0.5 for specification of the firms
group.
Diagram 1: Best prediction model of
financial crisis or bankruptcy obtained
from Non-Linear Genetic Algorithm
process
Given the above figure, the obtained model
can be presented as follows:
Y=(X16)2 + (X14 - X16 - ((X16 X19)
X14 X19)) + ((-X18)2 X14 X18)
In table 4, the number of the training
sample and the hold-out sample as well as
the number of errors in two samples are
separately presented for the bankrupt and
non-bankrupt firms in the Non-Linear
Genetic Algorithm model.

Vol.2 / No.1 / Winter 2012

International Journal of Finance, Accounting and Economics Studies / 19

+
*

+
^
X16

X14

*
X16

The Non-Linear Genetic Algorithm

succeeded in correct classification of firms
present in the training sample into groups
of bankrupt and non-bankrupt firms with
an overall accuracy of 91%, so as from
among 104 firms present in the training
set, 85 firms have been correctly classified.
Study of the results indicates that the Nonlinear Genetic Algorithm had an accuracy
of 86 percent in correct classification of
firms in the training set (from among 51
bankrupt firms in this set, 44 firms have
been correctly classified). In addition, this
model had an accuracy of 96 percent in
correct classification of non-bankrupt firms
in the training set (from among 53 nonbankrupt firms present in this set, 51 firms
have been correctly classified).
To examine generalizability and
reliability of the Linear Genetic Algorithm,
the model has been tested on the data of 40
firms present in the hold-our sample. The
firms placed in the hold-out set had no
interference in the models construction
process. Hence, they can be correctly used
to test the models external validity. The
Non-Linear Genetic Algorithm succeeded
in correct classification of firms present in

X19

*
X14

*
X16

X18

*
-1

X14

X18

X19

the hold-our sample into groups of

bankrupt and non-bankrupt firms with an
overall accuracy of 90%, so as from among
40 firms present in the hold-out set, 36
firms have been correctly classified. The
Non-linear Genetic Algorithm model had
an accuracy of 81 percent in correct
classification of firms in the hold-out set
(from among 21 firms in this set, 17
bankrupt firms have been correctly
classified). In addition, this model had an
accuracy of 100% in correct classification
of firms in the hold-out set (from among
19 non-bankrupt firms in this set, 19 firms
have been correctly classified). These
results indicate that the Non-linear Genetic
Algorithm in addition to producing
desirable results in prediction of firms
future state based on their financial
information is not biased in classification
of bankrupt or non-bankrupt firms and
creates balanced and reliable results.
Turbulence Matrix of the Non-linear
Genetic Algorithm prediction result is
shown in table 5.

Vol.2 / No.1 / Winter 2012

20 /

Investigating Financial Crisis Prediction Power using Neural Network

Table 4: Number of sample and errors in Non-Linear Genetic Algorithm
Training sample
Number of sample Number of error
53
2
51
2
104
9

Non-bankrupt
Bankrupt
Total

Hold-out sample
Number of sample Number of error
19
0
21
4
40
4

Table 5: Turbulence Matrix of Non-Linear Genetic Algorithm prediction result

Real
Bankrupt
Non-bankrupt

Predicted

Bankrupt
TP = 17
FP = 4

Non-bankrupt
FN = 0
TN = 19

Table 6: Results of Neural Network technique for the training sample

Non-bankrupt
Bankrupt
Total

Training sample
Hold-out sample
Number of sample Number of error Number of sample Number of error
53
15
19
6
51
12
21
6
104
27
40
12

Thus, the accuracy of Non-Linear

Genetic Algorithm model which is 90% in
the hold-out sample can be calculated as
follows:
Accuracy
of
Algorithm =

Non-Linear

Genetic

= 90%

7. Construction of Financial Crisis or

Bankruptcy Prediction model using
Neural Network
The used Neural Network is a fully
interconnected network in which the
learning algorithm Error-Back Propagation
(EBP) is utilized for the purpose of
training. The used Transformation
Function in each neuron of this network is
a sigmoid function as follows:
f(NET) = (1 + e

In this relation, NET is the weighted

sum of the neurons input variables from
the previous layer. Using this function, the
numerical output value will be zero or one.

Number of the training sample, number

of the hold-out sample, and number of
errors in the two samples for bankrupt and
non-bankrupt
firms
separately
are
presented in table 6.
Neural Network model succeeded in
correct classification of the existing firms
in the training sample into bankrupt and
non-bankrupt groups with an overall
accuracy of 74%, so as from among 104
companies present in the training set, 77
companies have been correctly classified.
Study of the results of this model
indicates that the Neural Network model in
correct classification of the bankrupt firms
in the training set has an accuracy of 76%
(from among 51 bankrupt firms in this set,
39 firms have been correctly classified). In
addition,
this
model
in
correct
classification of the non-bankrupt firms in
the training set has an accuracy of 72%
(from among 53 non-bankrupt firms in this
set, 38 firms have been correctly
classified).

Vol.2 / No.1 / Winter 2012

International Journal of Finance, Accounting and Economics Studies / 21

To examine generalizability and
reliability of the Neural Network model,
this model has been tested on the data
regarding the 40 firms present in the holdout sample. The firms placed in the holdout sample have had no part in the model
construction process. Hence, they can be
properly used to test the models external
validity.
The Neural Network model succeeded
in correct classification of the existing
firms in the hold-out sample into bankrupt
and non-bankrupt groups with an accuracy
of 70%, so as from among 40 firms present
in the hold-out set, 28 firms have been
correctly classified. The Neural Network
model in correct classification of bankrupt
firms in the hold-out set has an accuracy of
71% (from among 21 bankrupt firms in
this set, 15 firms have been correctly
classified). In addition, this model in
correct classification of non-bankrupt firms
in the hold-out set has an accuracy of 68%
(from among the 19 non-bankrupt firms in
this set, 13 firms have been correctly
classified).
These results indicate that the Neural
Network model in addition to generation of
desirable results in prediction of firms
future state using their financial
information, is not biased towards either
bankrupt firms or non-bankrupt firms and
produces balanced and reliable results.
As is observed, there are 40 firms in
total in the hold-out sample consisted of 19
non-bankrupt firms and 21 bankrupt firms,
and in the built model, from among 19
non-bankrupt firms 6 firms have been
erroneously predicted (i.e. error of second
type is equal to 6). In addition, in the
constructed model, from among the 21
bankrupt firms 6 firms have been
erroneously predicted (i.e. error of first
type is equal to 6). In sum, total number of
errors is equal to 12 firms whose

bankruptcy or non-bankruptcy has been

incorrectly predicted. The Turbulence
Matrix of the Neural Network model
prediction result is provided in table 7.
Table 7: Results of Neural Network
Technique for the hold-out sample
Real

TP = 15

Nonbankrupt
FN = 6

FP = 6

TN = 13

Bankrupt
Predicted

Bankrupt
Nonbankrupt

Thus, the accuracy of Neural Network

model which is 70% in the hold-out
sample can be calculated as follows:
Accuracy
of
Non-Linear
Genetic
Algorithm =
=

= 70%

8. Conclusion
To test the hypotheses of this research,
prediction accuracy of the generated
models in previous sections is compared
using Linear and Non-Linear Genetic
Algorithm and Neural Networks and
MacNemar Test. As was observed,
prediction accuracy of the models
generated by Linear Genetic Algorithm,
Non-Linear Genetic Algorithm, and Neural
Network was 80, 90, and 70 percent,
respectively. Now, the question arises as
whether these differences between
prediction accuracy of the constructed
models using these techniques are
significant or not? In this test, results of the
generated models are compared with each
other to decide about presence of any
significant difference between them.
Results of this test for presence of any
significant difference between results of
the models generated by Non-Linear
Genetic Algorithm and Neural Network
are provided in table 8. In general, results
of this research indicate that prediction of
financial crisis or bankruptcy is possible in

Vol.2 / No.1 / Winter 2012

22 /

Investigating Financial Crisis Prediction Power using Neural Network

the Iranian economic environment. In

addition, since this prediction is made
based on the financial information
available in the firms financial statements,
it can be per se an evidence for presence of
informative content of financial statement
for more optimal functioning of the capital
market. The findings of this study are
consistent with research results of Huang,
Tsai, Yen and Chang (2008), Hong and
Cheng (2008), Lin et al (2009), Min and
Jeong (2009), Min and Jeong (2008), Ravi
and Pramod (2008), Sun and Li (2008),
and Wu (2010).
Results of McNemar Test for NonLinear Genetic Algorithm and Neural
Network in table 4.12 indicate since the
significance level is greater than 5%
(0.059), there is no significant difference
between results of Non-Linear Genetic
Algorithm and Neural Network. Although

prediction power of Non-Linear Genetic

Algorithm (90%) is greater than that of
Neural Network (70%), this difference is
not statistically significant and accordingly
the second sub-hypothesis suggesting a
greater prediction power for the models
based on Neural Networks relative to NonLinear Genetic Algorithm is not
confirmed.
Although prediction power of NonLinear Genetic Algorithm is greater than
that of Linear Genetic Algorithm and
prediction power of the both is greater than
that of Neural Networks, yet these
differences are not statistically significant
and in sum, the main hypothesis of this
research suggesting a greater prediction
power for the models based on Neural
Networks
compared
to
Artificial
Intelligence techniques based on internal
analysis (Genetic Algorithm) is rejected.

Table 8: Results of McNemar Test for Linear Genetic Algorithm and Neural Network
techniques
Test Statisticsb
NN & GANL
NN

NN & GANL

GANL
0

58

11

23

52

144

Chi-Squarea

3.559

Asymp. Sig.

.059

a. Continuity Corrected
b. McNemar Test

9. Research Suggestions
Based on the results of this research the
following suggestions can be offered:
The investors are recommended to use
Non-linear Genetic Algorithm model as
well as Neural Network model in
evaluation of Iranian firms financial
state and in decision making regarding
their investment. Investors should note

that since in this research article 141 is

used for definition of bankrupt firms, it
does not lead to immediate dissolution
of the firms and suspension of their
activities.
The
Securities
and
Exchange
Organization is recommended to use
these models for admission of firms to
the stock exchange and for evaluation

Vol.2 / No.1 / Winter 2012

International Journal of Finance, Accounting and Economics Studies / 23

of firms operating on the stock
exchange.
Auditors are recommended to use these
models in their comments on continuity
of the audited firms operation.
One of the main limitations of this
research is unavailability of all the required
information for calculation of financial
ratios. If in the future this information is
made
available,
researchers
are
recommended by applying the financial
ratios,
particularly financial
ratios
regarding cash flows, to redesign a model
for prediction of financial crisis or
bankruptcy.
Due to the high efficiency and
effectiveness of Non-linear Genetic
Algorithm and Neural Network in solving
complex problems, researchers in the area
of finance are recommended to utilize
these models in other areas such as
prediction of price, share return, the used
indices, etc and compare their results with
those obtained from the current techniques.
References
1) Altman, E.I. (1968), Financial Ratios,
Discriminant Analysis and Prediction
of Corporate Bankruptcy, Journal of
Finance, 23, pp. 589609.
2) Beaver, W.H. (1966), "Financial Ratios
as Predictors of Failure", Journal of
Accounting Research 4, Empirical
Research in Accounting: Selected
studies, pp. 71-111.
3) Beaver, W. H., McNichols, M. F.,
Rhie, J. W., (2005), Have Financial
Statements Become Less Informative?
Evidence from Power of Financial
Ratios in Prediction of Bankruptcy,
Review of Accounting Studies, 10, pp.
93122.
4) Etemadi, H., Rostamy, A. & Dehkordi,
H. (2009), "A Genetic Programming

Model for Bankruptcy Prediction:

Empirical Evidence from Iran", Expert
Systems with Applications 36(2), pp.
31993207.
5) Farajzadeh
Dehkordi,
Hassan,
Etemadi, Hossein, & Rostami,
Aliasghar (2005), Application of
Genetic Algorithm in Bankruptcy
Prediction Modeling, Master Thesis
in Accounting, Faculty of Human
Sciences,
Teachers
Training
University, p.80
6) Faghih, Nazamoddin (2004), Genetic
Algorithm in Planning of Predictive
Inspections, Shiraz, Nassim-e-Hayat
Publications
7) Galvao, R.K. (2004), Ratio Selection
for Classification Models, Data
Mining and Knowledge Discovery, 8,
pp. 151170.
8) Goldberg, D. E. (1989), Genetic
Algorithms in Search, Optimization
and Machine Learning, New York:
Addison-Wesley.
9) Gordon, M.J. (1971), Towards a
Theory of Financial Distress, the
Journal of Finance, 36, pp. 47-56.
10) Haber,
J.
(2006),
Theoretical
Development of Bankruptcy Prediction
Variables, the Journal of Theoretical
Accounting Research, 2, pp. 82-101
11) Huang, S., Tsai, C. F., Yen, D. &
Cheng, Y. (2008), "A Hybrid Financial
Analysis Model for Business Failure
Prediction", Expert Systems with
Applications 35(3), pp. 10341040.
12) Hung, C., & Chen, J. (2009), "A
Selective Ensemble based on Expected
Probabilities
for
Bankruptcy
Prediction", Expert Systems with
Applications 36(3), pp. 52975303.
13) Kiarasi, Ava, Pourzamain, Zahra, and
Tavangar, Afsaneh (2009), Efficiency
Comparison of Logistic Regression
and
Multivariate
Discriminant

Vol.2 / No.1 / Winter 2012

24 /

Investigating Financial Crisis Prediction Power using Neural Network

Analysis Models in Detection of

Firms Financial Ability (Listed Firms
on the Stock Exchange), Master
thesis, Faculty of Economics and
Accounting, Islamic Azad University,
Tehran Central Branch
14) Lin, R., Wang, Y., Wu, C., et al.
(2009), "Developing a Business Failure
Prediction Model via RST, GRA and
CBR",
Expert
Systems
with
Applications 36(2), pp. 15931600.
15) Mansour,
Jahangir
(2000),
Commercial Law along with the
Check Act, and the Amended Bylaw
for Registration of Non-Commercial
Organizations, Didar Publication, 8th
edition
16) McKee, T.E., Lensberg, T. (2002),
Genetic Programming and Rough
Sets: a Hybrid Approach to
Bankruptcy Classification, European
Journal of Operational Research, 138,
pp. 436-51.
17) Min, J., & Jeong, C. (2009), "A Binary
Classification Method for Bankruptcy
Prediction", Expert Systems with
Applications 36(3), pp. 52565263.
18) Min, J.H., & Lee, Y.C. (2008), "A
Practical Approach to Credit Scoring",
Expert Systems with Applications
35(4), pp. 17621770.
19) Newton, G.W. (1998), Bankruptcy
Insolvency Accounting Practice and
Procedure, 1: Wiley, pp. 21-41
20) Nouraddin, Mostafa, Pourzamani,
Zahra & Keypour, Reza (2010), Study
and Comparison of Bankruptcy
Models Prediction Power (Understudy
Models based on Traditional Financial
Ratios, Genetic Algorithm, and Neural
Network), Master Thesis, Faculty of
Economics and Accounting, Islamic
Azad University, Tehran Central
Branch

21) Odom, M. & Sharda, R. (1993),

"Neural Network for Bankruptcy
Prediction, in: Trippi, Robert &
Turban, Efrain, Neural Network in
Finance and Investment: Using
Artificial Intelligence to Improve RealWorld
Performance",
Probus
Publishing Company, pp.177-185.
22) Ravi, V., & Pramodh, C. (2008),
"Threshold
Accepting
Trained
Principal Component Neural network
and
Feature
Subset
Selection:
Application to Bankruptcy Prediction
in Banks", Applied Soft Computing
8(4), pp. 15391548.
23) Saee, Reza & Fallahpour, Saeed
(2008), Application of Support Vector
Machine in Prediction of Firms
Financial Insolvency using Financial
Ratios,
Quarterly
Journal
of
Accounting and Audit Studies, 53: 1734
24) Serrano, Cinca, C. (1996-97), "FeedForward
Neural
Networks
in
Classification
of
Financial
Information", the European Journal of
Finance 3, pp. 183-202.
25) Shah, J.R., Murtaza, M.B. (2000), "A
Neural Network Based Clustering
Procedure for Bankruptcy Prediction",
American Business Review 18(2), pp.
80-86.
26) Shin, K., Lee, Y. (2002), A Genetic
Algorithm Application in Bankruptcy
Prediction Modeling, Expert Systems
with Applications, 23 (3), pp. 321-8.
27) Sun, J., & Li, H. (2008a), "Prediction
of Listed Companies Financial
Distress Based on Weighted Majority
Voting Combination of Multiple
Classifiers", Expert Systems with
Applications 35(3), pp. 818827.
28) Tam, K.Y., & Kiang, M.Y. (1992),
"Managerial Applications of Neural
Network: The case of Bank Failure

Vol.2 / No.1 / Winter 2012

International Journal of Finance, Accounting and Economics Studies / 25

Predictions", Management Science
38(7), pp. 926-947.
29) Tang, T. C., Chi, L. C. (2005), "
Neural Networks Analysis in Business
Failure
Prediction
of
Chinese
Importers:
A
between-Countries
Approach", Expert Systems with
Applications 29, pp. 244255.
30) The
Securities
and
Exchange
Organization of Tehran, the Stock
Exchange Council (September, 1999),
Collection of Laws and Bylaws on the
Stock Exchange, articles 75 and 76,
Dissolution of Companies, First edition
31) Tsai, C.F. (2009), "Feature Selection in
Bankruptcy Prediction", KnowledgeBased Systems 22, pp. 120127.
32) Whitaker, R., 1999. The Early Stays
of Financial Distress, Journal of
Economics and Finance, 23(2), pp.
122-133.
33) Wu, W.W. (2010), "Beyond Business
Failure Prediction", Expert Systems
with Applications 37(3), p. 23712376.
34) Zhang, G., HU, M. Y., Patuwo, B. E.,
& Indro, D. C. (1999), "Artificial
Neural Network in Bankruptcy
Prediction: General Framework and
Cross-Validation Analysis", European
Journal of Operational Research
116(1), pp. 16-32

Vol.2 / No.1 / Winter 2012